Are Higher-Volume Hospitals Really Better?

Numerous studies test whether increased provider volume is correlated with improved outcomes, but few of these studies justify their regression model or control for potential unobserved heterogeneity among providers. A literature search on this topic yielded 85 studies published between 2008 and 2013 that examine the volume-outcome effect; only nine of these studies employed panel-data methods, and only one relied solely on within-hospital variation in volume when estimating the volume-outcome effect. We use a large data set of cancer patients to examine alternative regression models and their implications for the volume-outcome relationship and health policy.

Our study investigates the relationship between hospital volume and in-hospital mortality for six cancer procedures (colectomy, esophagectomy, pancreatic resection, pneumonectomy, pulmonary lobectomy, and rectal resection). Using patient-level hospital discharge data from Florida, New Jersey, and New York from 2000 through 2011, we select all patients for whom the principal procedure code and at least one diagnosis code match the ICD-9 codes for the corresponding procedure and cancer. We then merge each data set with AHA hospital data and fit four regression models (all but the first using clustered standard errors): baseline logit, logit, hospital fixed effects, and hospital random effects.

Logit models are widely used in volume-outcome studies, but they yield inconsistent estimates if there are unobservable factors contributing to mortality variation across hospitals. Coefficient estimates in the fixed-effects model are consistent, but at the cost of ignoring between-hospital variation -- estimates are based solely on within-hospital variation. Coefficient estimates in the random-effects model use both within- and between-hospital variation and hence are more efficient. However, the random-effects model assumes that hospital-specific unobserved characteristics are independent of the other regressors; if the unobservable factors are correlated with volume, these estimates are at best inconsistent, and at worst misleading.

We find that the statistical significance of hospital volume depends on the regression model used. Regressions with both logit models and the random-effects model yield significant volume effects for all procedures except esophagectomy (the smallest data set), whereas fixed-effects regressions yield insignificant volume effects, despite the sizable percentage of variation in hospital volume coming from within-hospital variation (minimum 29.6%, lobectomy; maximum 40.8%, pneumonectomy). For three of the procedures (colectomy, pancreatic resection, and rectal resection), the Hausman test rejects the hypothesis that the random-effects estimator is consistent, thus favoring the fixed-effects model.

Previous findings of a volume-outcome relationship have been used to support recommendations for centralizing complex procedures, referring patients to high-volume hospitals, and continuing support of state certificate-of-need regulations. However, most clinical and empirical research examining the volume-outcome relationship employs the logit regression framework without fixed effects. By ignoring institution-specific effects, these studies may overestimate the effect of volume in reducing mortality. Careful selection and verification of estimation methods must occur before further policy recommendations are made.